John F. Kennedy once said, “Man is still the most extraordinary computer of all.” While that’s undeniably still true, can a machine learn from the information it takes in and then make better decisions, just like a human?
Thanks to a group of Northrop Grumman employees engaged in what’s known as “deep learning,” this area of artificial intelligence is taking shape to a fascinating degree.
Deep learning is a class of machine learning algorithms that learn features directly from data to achieve some task. (Algorithms are a process or set of rules to be followed in calculations or other problem-solving operations, in most cases by a computer.)
The Cats and Dogs of Deep Learning
In this day of Big Data, analysts and warfighters have access to a mounting volume of data coming from various sensors. A network using deep learning technology can help bring order to the avalanche, and find the critical items that decision-makers need to take action on. It can enable fast solutions or recommendations to enhance human decision making when speed is paramount.
“For example, if I give the deep network pictures of dogs and cats, and tell the network which are dogs and which are cats, it can learn the features that distinguish cats from dogs,” says Dr. Will Chambers, a Northrop Grumman research and development engineer with expertise in artificial intelligence. “So the next time it sees a dog, it can quickly identify that it’s a dog.”
In the traditional approach, you would have to define the features, telling the algorithm what dogs and cats look like. With that approach, the algorithm’s performance is limited. “So you want an algorithm that can generalize beyond the data that you’ve given it, like people do,” says Chambers, whose team is based in Aurora, Colo. Thus, the primary advantage to deep learning networks is their effectiveness with generalizing.
The Three Flavors of Deep Learning
There are three types of deep (or machine) learning:
• Supervised learning, in which you provide the images and labels (tell the machine which are dogs and cats)
• Unsupervised learning, or clustering algorithms, in which you just give the data — like pictures of dogs and cats with no labels — and let the machine sort them out
• Reinforcement learning, in which you try to train an agent to do something — to map an observation to an action, by providing the agent rewards
Chambers says he got involved in deep learning when he started seeing the results. “Those results, in terms of classification problems with images, were blowing away any of the prior traditional machine learning approaches based on human-engineered features,” he says.
“Deep learning is shaping industry,” says Chambers. “It’s satisfying to see that our customer community is finally embracing it as an applicable technology, whereas a couple years ago it was still viewed with skepticism.”
Charlie Parkinson, a program manager of Aurora’s Autonomous Intelligence and Robotics lab, says it’s been interesting to see how we can apply learning algorithms to traditional radio frequency spaces to ultimately provide better tools and intelligence to our customers.
“It’s opening up new doors that our customers are interested in hearing about,” Parkinson says. “It’s also been a great opportunity to get some junior engineers up to speed on bringing the latest technologies into our domains.”